From 19199cd38d826ad146ee9e4a0cdfed78e347ff10 Mon Sep 17 00:00:00 2001
From: Teknium <127238744+teknium1@users.noreply.github.com>
Date: Mon, 13 Apr 2026 23:11:13 -0700
Subject: [PATCH] fix: clamp 'minimal' reasoning effort to 'low' on Responses
 API (#9429)

GPT-5.4 supports none/low/medium/high/xhigh but not 'minimal'.
Users may configure 'minimal' via OpenRouter conventions, which would
cause a 400 on native OpenAI. Clamp to 'low' in the codex_responses
path before sending.
---
 run_agent.py                                  |  6 ++
 .../test_run_agent_codex_responses.py         | 63 +++++++++++++++++++
 2 files changed, 69 insertions(+)

diff --git a/run_agent.py b/run_agent.py
index 592253464..626951b27 100644
--- a/run_agent.py
+++ b/run_agent.py
@@ -6143,6 +6143,12 @@ class AIAgent:
                 elif self.reasoning_config.get("effort"):
                     reasoning_effort = self.reasoning_config["effort"]
 
+            # Clamp effort levels not supported by the Responses API model.
+            # GPT-5.4 supports none/low/medium/high/xhigh but not "minimal".
+            # "minimal" is valid on OpenRouter and GPT-5 but fails on 5.2/5.4.
+            _effort_clamp = {"minimal": "low"}
+            reasoning_effort = _effort_clamp.get(reasoning_effort, reasoning_effort)
+
             kwargs = {
                 "model": self.model,
                 "instructions": instructions,
diff --git a/tests/run_agent/test_run_agent_codex_responses.py b/tests/run_agent/test_run_agent_codex_responses.py
index 0fca9e4df..785d85886 100644
--- a/tests/run_agent/test_run_agent_codex_responses.py
+++ b/tests/run_agent/test_run_agent_codex_responses.py
@@ -287,6 +287,69 @@ def test_build_api_kwargs_codex(monkeypatch):
     assert "extra_body" not in kwargs
 
 
+def test_build_api_kwargs_codex_clamps_minimal_effort(monkeypatch):
+    """'minimal' reasoning effort is clamped to 'low' on the Responses API.
+
+    GPT-5.4 supports none/low/medium/high/xhigh but NOT 'minimal'.
+    Users may configure 'minimal' via OpenRouter conventions, so the Codex
+    Responses path must clamp it to the nearest supported level.
+    """
+    _patch_agent_bootstrap(monkeypatch)
+
+    agent = run_agent.AIAgent(
+        model="gpt-5-codex",
+        base_url="https://chatgpt.com/backend-api/codex",
+        api_key="codex-token",
+        quiet_mode=True,
+        max_iterations=4,
+        skip_context_files=True,
+        skip_memory=True,
+        reasoning_config={"enabled": True, "effort": "minimal"},
+    )
+    agent._cleanup_task_resources = lambda task_id: None
+    agent._persist_session = lambda messages, history=None: None
+    agent._save_trajectory = lambda messages, user_message, completed: None
+    agent._save_session_log = lambda messages: None
+
+    kwargs = agent._build_api_kwargs(
+        [
+            {"role": "system", "content": "You are Hermes."},
+            {"role": "user", "content": "Ping"},
+        ]
+    )
+
+    assert kwargs["reasoning"]["effort"] == "low"
+
+
+def test_build_api_kwargs_codex_preserves_supported_efforts(monkeypatch):
+    """Effort levels natively supported by the Responses API pass through unchanged."""
+    _patch_agent_bootstrap(monkeypatch)
+
+    for effort in ("low", "medium", "high", "xhigh"):
+        agent = run_agent.AIAgent(
+            model="gpt-5-codex",
+            base_url="https://chatgpt.com/backend-api/codex",
+            api_key="codex-token",
+            quiet_mode=True,
+            max_iterations=4,
+            skip_context_files=True,
+            skip_memory=True,
+            reasoning_config={"enabled": True, "effort": effort},
+        )
+        agent._cleanup_task_resources = lambda task_id: None
+        agent._persist_session = lambda messages, history=None: None
+        agent._save_trajectory = lambda messages, user_message, completed: None
+        agent._save_session_log = lambda messages: None
+
+        kwargs = agent._build_api_kwargs(
+            [
+                {"role": "system", "content": "sys"},
+                {"role": "user", "content": "hi"},
+            ]
+        )
+        assert kwargs["reasoning"]["effort"] == effort, f"{effort} should pass through unchanged"
+
+
 def test_build_api_kwargs_copilot_responses_omits_openai_only_fields(monkeypatch):
     agent = _build_copilot_agent(monkeypatch)
     kwargs = agent._build_api_kwargs([{"role": "user", "content": "hi"}])